Linux 内核 TCP 参数调优

配置项位置说明：

net.core.{option} = /proc/sys/net/core/{option}
net.ipv4.{option} = /proc/sys/net/ipv4/{option}

如何修改参数：

#临时设置-系统参数文件：
echo "{SOME_VALUE}" > /proc/sys/net/ipv4/tcp_xxx

#临时设置-执行命令:
systemctl -w net.core.{option} = {value}

#永久生效-需要修改: /etc/sysctl.conf
sysctl -p

有些参数的生效需要刷新路由信息: sysctl -w net.ipv4.route.flush=1

性能验证：

dd if=/dev/urandom of=sample.txt bs=1M count=1024 iflag=fullblock
scp sample.txt [email protected]:/some/remote/directory

查看统计:

ls /sys/class/net/eth0/statistics/

初始设置

#设置可分配端口范围
net.ipv4.ip_local_port_range = 9000 65000
#内核从NIC收到包后,交由协议栈(如IP、TCP)处理之前的缓冲队列长度
net.core.netdev_max_backlog = 16384

窗口 | 缓冲区大小

net.ipv4.tcp_mem        = 327680    327680      16777216
net.ipv4.tcp_wmem       = 4096      65536       16777216
net.ipv4.tcp_rmem       = 4096      327680      16777216
net.core.rmem_default   = 327680
net.core.wmem_default   = 65536
net.core.rmem_max       = 16777216
net.core.wmem_max       = 16777216

# 设置网卡队列大小
ifconfig eth0 txqueuelen 1000

建立连接

net.ipv4.tcp_fastopen           = 3
net.ipv4.tcp_syn_retries        = 2
net.ipv4.tcp_synack_retries     = 2

# TCP Backlog队列实际值 = min(somaxconn, tcp_max_syn_backlog, 应用层设置的backlog)
# TCP accept队列实际值 = min(somaxconn, 应用层设置的backlog)
net.core.somaxconn              = 16384 #全连接队列
net.ipv4.tcp_max_syn_backlog    = 16384 #半连接队列
# 0 表示如果 accept 队列满了，那么扔掉ACK ； 1表示如果accept队列满了，server发送一个 RST
net.ipv4.tcp_abort_on_overflow  = 1
# syncookies有三个值： 0表示关闭该功能； 1表示仅当半连接队列放不下时再启用它； 2表示无条件开启功能；
net.ipv4.tcp_syncookies         = 1

连接使用

net.ipv4.tcp_keepalive_time     = 600
net.ipv4.tcp_keepalive_probes   = 3
net.ipv4.tcp_keepalive_intvl    = 15
net.ipv4.tcp_retries1           = 3
net.ipv4.tcp_retries2           = 5
net.ipv4.tcp_no_delay_ack       = 1
net.ipv4.tcp_low_latency        = 1

拥塞控制

#设置拥塞窗口大小
ip route | while read p; do ip route change $p initcwnd 10 initrwnd 10; done
net.ipv4.tcp_window_scaling = 1

#空间后不进入慢启动
net.ipv4.tcp_slow_start_after_idle = 0

#允许最大重排数量
net.ipv4.tcp_reordering = 8

#拥塞算法
net.ipv4.tcp_available_congestion_control = cubic reno bbr
net.ipv4.tcp_congestion_control = bbr

#重传策略
net.ipv4.tcp_sack   = 1
net.ipv4.tcp_fack   = 1 #新版内核不再使用
net.ipv4.tcp_dsack  = 1

连接回收

net.ipv4.tcp_tw_reuse       = 1
net.ipv4.tcp_tw_recycle     = 1
net.ipv4.tcp_timestamps     = 0

net.ipv4.tcp_fin_timeout    = 2
net.ipv4.tcp_max_tw_buckets = 6000